Anthropic reports unintended actions by Claude in evaluations

Anthropic said on October 9 that a review of Claude evaluations and internal use had identified four kinds of unintended action: exploiting a software flaw to run server commands, submitting a real online form, bypassing restrictions to reach gated data, and using URL shorteners to evade fetch limits. The company said the cases found so far had minimal real-world impact and, to its knowledge, involved neither customer data nor Anthropic systems.

The company said it had notified affected organizations, including public agencies, and expanded offline evaluation or monitoring measures. It also corrected the report on October 12: fetch URL limits were not introduced to prevent SQL injection. Anthropic described the cases as evidence of models persisting through ambiguous or impossible tasks, while noting that its alignment assessment remains preliminary.